Better speech recognition for Chinese dialects through unified romanization
Summary and headline written by AI from the source article. How we work
Researchers developed a new system for romanizing Sinitic languages, languages related to Chinese, to improve speech recognition across dialects.
The Sinitic Romanization Ecosystem aims to standardise how different Chinese languages are written using the Roman alphabet. Currently, inconsistent romanization methods hinder the development of speech technology for less-studied dialects. The team’s framework prioritises phonetic and historical links between languages. It uses consistent symbols for similar sounds and cognates, words with shared origins, while sticking to basic Latin letters.
They created paired romanization schemes, CantRomZJ1 for Cantonese and MandRomZJ1 for Mandarin, and applied the framework to other Sinitic languages including Meixian Hakka, Shanghai Wu, and Nanjing Jianghuai Mandarin. These languages represent diverse regions and linguistic groups within China. To support practical use, the researchers built open-source tools for storing, converting, and parsing romanized text. They also designed the system to work with input methods, how users type on computers, and to construct dictionaries.
Testing the system with Meta’s Massively Multilingual Speech dataset showed a significant improvement in Cantonese speech recognition. The new romanization reduced Cantonese Word Error Rate by 7.80% and Character Error Rate by 10.61% compared to existing Pinyin and Jyutping systems. This work demonstrates that aligning romanization across Sinitic languages can boost speech technology performance for under-resourced dialects, potentially preserving linguistic diversity and improving accessibility.
